People call it thermal throttling, Intel calls it Turboboost.
They are different things. Thermal throttling is lowering the clock speed below base speed because it is too hot, caused by a manufacturer not following Intel’s guidelines on the amount of cooling required and it has a negative affect on performance. Turbo boost is running cores at a higher frequency when there is thermal capacity to do so, i.e. when not all cores are in use at the same time to boost performance.
It's not really based on failure to follow Intel guidelines. Intel has itself published info on when throttling is expected -- because it's part of the chip design when under a max load and the processor can't keep up. It happens; intensive gaming for example. In the real world thermal constraints cannot be avoided in our modern age of high-itensity computing. It's simply a matter of when, not if.
Agreed -Thermal throttling is a safeguard to protect the chip from damage due to overheating. How soon it happens depends on the design and efficiency of the heatsink/cooling mechanism designed by the computer manufacturer. Intel controls the chip design and processing power per watt. The manufacturers control the rest
dewme said: The point I am making is in regard to any "Moore's Law" type of expectations around single threaded performance, which are whacky. Your point about the potential 20% raw improvements with Intel's Cove cores reinforces my point exactly. To classify Intel as "rubbish" is unwarranted, especially in terms of single threaded performance in high-end desktop platforms.
Your expectations for single thread performance are too pessimistic. Perhaps aderutter should not have mentioned anything about Moore's Law, but the 20% single core performance gain across 5 to 6 years is primarily the result of Intel not being about to fab desktop processors on their 10 nm fab. If they were able to, single core performance improvements from 2015 to 2020 would be in the 50% to 70% range.
Tiger Lake GB5 scores are hovering around 1400 to 1600 single core points. They could have had that last year if 10nm was working, and they wouldn't need to backport Willow Cove to 14nm for Rocket Lake to maintain a semblance of parity with the mobile chips. Willow Cove on 10nm at desktop TDPs could hit 1600 to 1800 GB5 single core points.
I would describe Intel's last two generations of desktop processors as rubbish, figurative turds. They have turbo power consumption levels of 150 to 250 W and added 125 W SKU line! All that for not much performance gain. That's nothing for Intel to be proud of, and I bet Intel OEMs aren't all that happy about it either.
Hmmm... Perhaps. Or perhaps Intel focused on additional cores for improving multi-threaded performance rather than focusing on single thread performance. I suspect they could build a perky little single core non-hyperthreaded CPU which would provide significant improvements in the much vaunted single thread performance. Unfortunately, the multi-thread performance would be worse than the single thread performance unless we go back to the multi-cpu motherboards.
This is not to say that a single cpu / single thread CPU would be 500% faster than a similar cpu from 6 years ago, but I would certainly expect Intel could achieve more than a 20% improvement. Just imagine the amount of L1 cache they could cram into one of those things!
dewme said: The point I am making is in regard to any "Moore's Law" type of expectations around single threaded performance, which are whacky. Your point about the potential 20% raw improvements with Intel's Cove cores reinforces my point exactly. To classify Intel as "rubbish" is unwarranted, especially in terms of single threaded performance in high-end desktop platforms.
Your expectations for single thread performance are too pessimistic. Perhaps aderutter should not have mentioned anything about Moore's Law, but the 20% single core performance gain across 5 to 6 years is primarily the result of Intel not being about to fab desktop processors on their 10 nm fab. If they were able to, single core performance improvements from 2015 to 2020 would be in the 50% to 70% range.
Tiger Lake GB5 scores are hovering around 1400 to 1600 single core points. They could have had that last year if 10nm was working, and they wouldn't need to backport Willow Cove to 14nm for Rocket Lake to maintain a semblance of parity with the mobile chips. Willow Cove on 10nm at desktop TDPs could hit 1600 to 1800 GB5 single core points.
I would describe Intel's last two generations of desktop processors as rubbish, figurative turds. They have turbo power consumption levels of 150 to 250 W and added 125 W SKU line! All that for not much performance gain. That's nothing for Intel to be proud of, and I bet Intel OEMs aren't all that happy about it either.
Hmmm... Perhaps. Or perhaps Intel focused on additional cores for improving multi-threaded performance rather than focusing on single thread performance. I suspect they could build a perky little single core non-hyperthreaded CPU which would provide significant improvements in the much vaunted single thread performance. Unfortunately, the multi-thread performance would be worse than the single thread performance unless we go back to the multi-cpu motherboards.
This is not to say that a single cpu / single thread CPU would be 500% faster than a similar cpu from 6 years ago, but I would certainly expect Intel could achieve more than a 20% improvement. Just imagine the amount of L1 cache they could cram into one of those things!
No, no. It's a lot worse than what you are thinking for Intel. They have been hamstrung on both number of cores and size of single cores for the better part of 3 years now, at least. Both require more transistors and that can't do either without making the chips too expensive for the consumer market.
They've been selling 8 to 28 core single socket Skylake Xeons for a long time now. A lot of them are defeatured and sold as rebranded Core X, but these come with $1000 to $2000 pice tags. They know how to do it very well. But, they simply can't go and take a 16 core Xeon design, put a UHD 630 in it, and ship it as a $600 Core i9. They'd lose money.
Likewise, higher performance single core performance requires more transistors. This makes it more expensive to fab on 14 nm and something would have to be left out. The Rocket Lake processors are likely to drop the 10-core SKU because they wouldn't be about to sell it at consumer prices.
This is also the reason why 14nm iGPUs have been poor performers. They can't afford the transistors.
So, they did the only thing they could do the last 3 years. Refine the fab to have incrementally have better yields on 4, 6, 8 and now 10 core Core designs, and incrementally increase clock rates for single core performance. This came at the expense of these insane turbo power levels. Really, figurative shit.
Tiger Lake on 10nm is a really nice mobile chip. It's the 10nm that really enables it niceness. The first real improvement in a long time. Would have been nice to see a Macbook based on it, last year that is. Apple is going to match it at half the Watts in a few months.
Comments
This is not to say that a single cpu / single thread CPU would be 500% faster than a similar cpu from 6 years ago, but I would certainly expect Intel could achieve more than a 20% improvement. Just imagine the amount of L1 cache they could cram into one of those things!
They've been selling 8 to 28 core single socket Skylake Xeons for a long time now. A lot of them are defeatured and sold as rebranded Core X, but these come with $1000 to $2000 pice tags. They know how to do it very well. But, they simply can't go and take a 16 core Xeon design, put a UHD 630 in it, and ship it as a $600 Core i9. They'd lose money.
Likewise, higher performance single core performance requires more transistors. This makes it more expensive to fab on 14 nm and something would have to be left out. The Rocket Lake processors are likely to drop the 10-core SKU because they wouldn't be about to sell it at consumer prices.
This is also the reason why 14nm iGPUs have been poor performers. They can't afford the transistors.
So, they did the only thing they could do the last 3 years. Refine the fab to have incrementally have better yields on 4, 6, 8 and now 10 core Core designs, and incrementally increase clock rates for single core performance. This came at the expense of these insane turbo power levels. Really, figurative shit.
Tiger Lake on 10nm is a really nice mobile chip. It's the 10nm that really enables it niceness. The first real improvement in a long time. Would have been nice to see a Macbook based on it, last year that is. Apple is going to match it at half the Watts in a few months.